Intel ISPC 1.30 Adds AMX Support, Boosts Performance on Intel and ARM Platforms

Intel has released version 1.30 of its ISPC (Implicit SPMD Program Compiler), adding support for Advanced Matrix Extensions (AMX) to the standard library. This update brings significant performance improvements for matrix operations on Intel's latest CPU architectures while also addressing performance regressions on Apple Silicon and other ARM platforms.

AMX Support for Matrix Operations

The headline feature in ISPC 1.30 is the addition of AMX support to the standard library. Intel's Advanced Matrix Extensions are designed to accelerate matrix multiplication operations, which are critical for machine learning workloads, scientific computing, and graphics processing.

With this release, developers can now leverage AMX instructions through a new amx.isph header file. The AMX implementation in ISPC works on capable hardware including:

• Sapphire Rapids CPUs
• Granite Rapids CPUs
• Upcoming Diamond Rapids processors

This integration allows developers using ISPC to automatically benefit from AMX acceleration without needing to write low-level assembly code or complex intrinsics. The SPMD programming model of ISPC makes it particularly well-suited for these types of data-parallel operations.

Performance Optimizations

Beyond AMX support, Intel has implemented several performance optimizations in this release:

Backward Memory Access Patterns: The compiler now generates more efficient code for workloads that access memory in reverse order, which is common in certain algorithmic patterns and can significantly impact cache performance.

Un-deprecation of SSE2 Targets: Based on customer feedback, Intel has un-deprecated the sse2-i32x4 and sse2-i32x8 targets. These 32-bit integer vector targets remain important for developers targeting older hardware or specific performance profiles.

Bug Fixes and Platform Improvements

One of the most significant bug fixes addresses a performance regression that affected Apple Silicon and other ARM platforms. Some workloads were experiencing up to 30% slowdowns when running on these architectures. This regression has been resolved in ISPC 1.30, restoring expected performance levels.

This fix is particularly important as Apple Silicon continues to gain market share in both consumer and professional computing segments. Developers using ISPC for cross-platform development can now expect consistent performance across x86 and ARM architectures.

Availability and Documentation

Intel ISPC 1.30 is available for download through GitHub, where developers can also find detailed release notes and documentation. The project continues to evolve as a powerful tool for developers looking to write SPMD code that can target both Intel CPUs and GPUs efficiently.

For developers working on performance-critical applications involving matrix operations or seeking to optimize their code for Intel's latest CPU architectures, ISPC 1.30 represents a meaningful update that combines new capabilities with improved reliability across platforms.

The addition of AMX support demonstrates Intel's commitment to providing developers with tools that can fully leverage their hardware capabilities, while the cross-platform performance improvements ensure ISPC remains a viable choice for heterogeneous computing environments.